# 4300 stars, 300 downloads: Folkbench is a billboard for an API-relay storefront site, and the stars are for show · GitHub open-source project

> Source: https://stringers.io/en/github/gh_Folkbench__folkbench.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> The stars are most likely bought: the high count got it auto-reposted by trendshift, SkillsLLM, LinkedIn repost accounts, and tutorial sites, driving traffic to a website carrying sponsor slots and affiliate links. Real users are estimated at only a few hundred.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

Folkbench/folkbench · Folkbench

# 4300 stars, 300 downloads: Folkbench is a billboard for an API-relay storefront site, and the stars are for show

**What it is** A site that ranks AI API relay stations by uptime and price, plus a desktop tool that writes your chosen route into Claude Code / Codex config with one click.

**Why it blew up** The stars are most likely bought: high star counts got it picked up by trendshift, SkillsLLM, LinkedIn repost accounts, and tutorial sites, all of which funnel traffic to a site with sponsor slots and referral links. Real users are estimated at only a few hundred.

**D** **How you could approach it** Don't copy the star-buying, and don't build another switcher. Fill what it didn't: open-source, reproducible, free relay-station verification and uptime checks, or a price-comparison MCP for agents. Post the results on V2EX and linux.do to earn real stars.

Grade D: its stars came mainly from buying them, so following that path means copying a tactic GitHub cleans up. Switchers already have cc-switch with 150k stars, and the evaluation space has relayAPI, Grandet, OkkMax, and others. The adjacent verification niche has potential, but it isn't closely tied to Folkbench itself. Word on the street

Data (updated daily) · 2026-10-02 repos indexed · 3,762 stars · 7 days +-317

GitHub · stars

Data 2026-10-10 → 2026-11-08 (30 days, 30 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

On 2026 Oct 16 afternoon, a post appeared in the "Share your creations" section of V2EX: "Tired of switching routes and remotely setting up Claude Code for friends, so I hacked together a desktop tool, Folkbench Switch." The poster said they had an official account but kept hitting "banned constantly" and "Selected model is at capacity," so they bought several third-party APIs as backups, and had to stop midway to switch whenever one 502 went down. When friends asked how to configure Claude Code, they walked them through it with "screenshot after screenshot" on WeChat. The post got 4 replies, and the first asked: "Is this implemented by calling cc-switch?"

That's the repo, which went from creation to 4322 stars in 35 days. Then in Nov 7, it lost 554 stars in a single day. This piece covers how those stars were obtained, how the business behind it makes money, and, if you want your open-source project to get stars, what's worth learning here and what to avoid.

### What it is

Folkbench/folkbench, written in Rust, described as "AI API relay station evaluation rankings and switching tool." It's really two things:

- **Website folkbench.com** : lists AI API relay stations by model (48 models), with uptime, input/output prices, and sample counts. It has an "API quality check" where you enter your own key to self-test, plus a blog and a relay station directory, in Chinese, English, Japanese, and Spanish. The domain was registered in Jun 29, and the first blog post went live in Sep 17, before the repo.

- **Desktop client Folkbench Switch in the repo** (Tauri + Rust): pick a route and, with one click, write the Base URL and Key into the config for Claude Code, Codex, Gemini CLI, Aider, OpenCode, and Kimi CLI.

Open the README and the first screen isn't a feature list; it's a sponsor section: four relay stations, Modelflare, CheapAI, IKunCode, and FluxLane. Their links go through folkbench.com/go/ redirects, and comments keep the raw partner= and aff= referral codes. Next to them is a line reading "top sponsor slots reserved," plus a self-serve signup page, "become a sponsor" at /vendor/login (11-06 reads).

### How the stars grew

| Date | What happened | Cumulative stars |
|---|---|---|
| 10-02 | The GitHub org and repo were created the same day (6 minutes apart). | — |
| 10-11 | Switch beta.1 and beta.2 released. | — |
| 10-16 | V2EX self-promo post, 4 replies. | — |
| 10-17 |  | 1235 |
| 10-25 | Single-day +561, the biggest day. | 3562 |
| From ¥10-26 | +50 to +90 per day |  |
| 11-05 | Peak | 4322, 21 forks. |
| 11-07 | One day: -554 stars, -3 forks. | 3758, 18 forks. |

Sources: GitHub daily records, GitHub API (11-07 check still shows 3758 stars, 18 forks). The curve is stepped: from 10-12 to 10-14 over three days, about +760; from 10-18 to 10-20 over three days, about +900; from 10-22 to 10-24 over three days, about +1300, with each stretch falling back to a few dozen. A real viral post usually shows one peak that decays gradually, not this shape.

### How many of these stars are real

Several numbers don't match:

- **Downloads** : Switch has 6 versions in total. The installers and attachments together have 317 downloads (GitHub Releases, 11-07), and the most-downloaded, beta.5, has 119. 4300 stars against 300 downloads means a dozen-plus stars per download. For a desktop tool repo, you'd expect a decent share of people who starred it to install it.

- **Forks** : at the peak of 4322 stars there were only 21 forks, 0.5%. In the same circle, zzsting88/relayAPI has 5450 stars and 174 forks, and hanzohanzhe/Grandet has 1143 stars and 125 forks (11%).

- **Charts** : it never made GitHub's official trending for a single day. It only appeared on trendshift's own Rust language chart (peak position #2, 10-22).

- **People** : the org has only this one repo. The main committer's account was registered in Oct 10 and has 0 public repos. Almost no one discusses it on social media: searching "folkbench" on X over the last 60 days turns up nothing, and there's nothing substantial on HN, Reddit, or linux.do.

- **Star loss** : 11-05 and 11-06 star gains dropped to zero, then 11-07 saw a one-time loss of 554 stars and 3 forks. That's the most common pattern when GitHub purges star-farmed accounts.

So our judgment: **most of the stars were bought (inferred)** . We didn't examine each starring account individually, so this isn't 100% confirmed. But the downloads, forks, trending history, and star loss all point the same direction. Real users are estimated in the hundreds.

### What the star-buying bought

Stars alone are worth nothing; what matters is that others use the count as a headline. High star counts automatically placed it in a chain of aggregation and repost channels: trendshift's language chart, SkillsLLM ("4.3k★"), agentskillshub, LinkedIn "trending GitHub projects this week" repost accounts ("1k stars in under two weeks"), an Instagram reel "Why developers love Folkbench," an intro article on explainx 10-18, a tutorial site post "3.8K-star relay station benchmark … step-by-step deployment," and a Zhihu column "Daily Good Software Recommendations 2026-10-19." Almost all of this secondhand content sells the star count, and readers who click through land on folkbench.com.

A few other channels were real:

- **Chinese developer communities** : a self-promo post on V2EX, with QQ and Telegram group links in the README.

- **SEO** : pinyin-slug landing pages /zhongzhuanzhan-ceping (relay station evaluation, 24 stations), /zhongzhuanzhan-na-ge-hao (which one is good), /gpt-codex-zhongzhuanzhan-ceping, with a ranking page for each model. A long English post on dev.to, and every link in the README carries utm tags.

- **Relay stations come to them** : self-serve submissions via /vendor/login, and issues like "[Listing request] Twinkle Model — source relay, GPT 1 off" have appeared in the repo.

### Where the money comes from

GitHub usually doesn't talk about money, but this repo exists for money, so to be clear: the README sponsor slots and the referral codes behind the /go redirects are the site's revenue. The relay stations being evaluated are also its customers, so the rankings are not clean. On the GPT 6.1 Sol chart, sponsor Modelflare has 93.75% uptime and ranks first, ahead of several stations at 100%. Sponsor FluxLane has only 18 samples yet makes the top eight (11-06 reads). Revenue isn't public, and its scale is unknown.

The demand is real. Claude Code and Codex are used heavily in China through relay stations, and rug pulls, model downgrades, and model swaps happen constantly. Forums keep asking "Is there a reliable relay station evaluation?" Folkbench is built on a real need, but its stars didn't come from that need.

### Who it's competing with

In the same circle, there are plenty of open-source projects that came earlier and are more solid:

| Project | Scale (11-06) | Compared with Folkbench |
|---|---|---|
| farion1231/cc-switch | 152k stars, 10k forks | The de facto standard for switching, pure tool, no rankings for sale; Folkbench Switch is its stripped-down version plus referral links |
| zzsting88/relayAPI | 5450 stars, 174 forks | Same play: relay-station recommendations plus referral commissions. March had it first, and its detection code is closed source |
| qixing-jk/all-api-hub | 5049 stars | Browser extension that manages balances, check-ins and price comparisons across multiple relay accounts |
| panxunying/ai-coding-welfare | 1283 stars, 84 forks | Collection of free / charity relay stations plus a one-click config script |
| hanzohanzhe/Grandet | 1143 stars, 125 forks | A Skill + MCP for agents: compares 1800 official, inference and relay prices across vendors. October built it |
| prehisle/relay-pulse | 1109 stars | Self-hosted relay availability monitor |
| OkkMax (fanbidog/okkmax-web) | 201 stars | Focuses on "purity testing": uses Anthropic's thinking signatures to catch model swaps. Declares it takes no money that affects rankings |
| Chenking-maker/api-rank, fanbidog/ai-relay-rank | 27 stars, 5 stars | An open-source "live test leaderboard" with no stars |

On the site side there's also a string of review sites: aiapipk.com, RouterHubs, APIRanking, llmtest.cn and others. Apparently built specifically to copy Folkbench, I haven't seen it.

### If it were you

**One-line verdict:** Folkbench's stars are mostly bought. Copying it means copying a star-faker's playbook. The repost pickup is real, but GitHub's cleanup can strip over 10% in one go, and a download-to-fork ratio gives it away. Not worth copying. The projects next to it that actually earned their stars are the ones to study.

**Who it's for:** Chinese-speaking developers who use relay stations daily, have been burned by shutdowns and model swaps, and can write some Rust / TS / Python. This crowd is urgent and eager to share. A tool that really solves the problem can earn a normal fork ratio (Grandet got 1100 stars and 11% forks in one month).

**Who it's not for:** Anyone trying to quickly build a name through star-farming or a sponsor banner at the top of the README. If you get cleaned up once, both the project and your account lose reputation, and reputation is the one thing open source lets you build up.

**How to do it:**

- Don't build a switcher. cc-switch has 150k stars; a weaker clone of it would just be "cc-switch-lite," and the first reply on V2EX would ask whether it's a fork of cc-switch.

- Do what Folkbench failed to do: **reproducible, no-money testing** . An open-source CLI or GitHub Action that runs with your own key to measure uptime, latency, and whether the model was swapped (the thinking-signature verification approach from OkkMax). Results are generated locally and can be shared as screenshots. Public detection code is the selling point: relayAPI closed its detection, and Folkbench's rankings are mixed with sponsorships.

- Or build the agent-facing version: follow Grandet's path and package "which vendor works now and what it costs" as an MCP / Skill so Claude Code can look it up itself. Only October started on this kind of thing.

- Or change ecosystems: these projects are almost all Chinese-language. English and Japanese communities also use plenty of third-party resellers beyond OpenRouter, but there's no widely recognized open-source "verification" tool yet (inference, not a formal survey).

**How to get users:** Share on V2EX's creation board, linux.do, and nodeseek. A post with a result image, "I tested 10 vendors, these 3 are swapping models," works far better than describing features. Publish a public test report in the README every month, and aggregator sites and repost accounts will pick it up on their own. Don't put referral links at the top of the README; this crowd is very sensitive to "shilling."

**When to launch:** No rush, but it's most newsworthy when models update fast: the week or two after each new model release, or when relay stations raise prices or swap models en masse, is the right time to publish results.

**How you'd know you were wrong:** If Folkbench keeps gaining dozens of stars a day after losing 554, and Switch downloads run into the thousands, then its users are far more numerous than we estimate, and the bundle of "leaderboard + one-click switch" has real demand. Conversely, if the open-source detection projects (OkkMax, relay-pulse) are still at a few hundred stars months later, this crowd just wants an answer to "which one should I use," doesn't care whether detection is open, and building an open-source detection tool won't earn many stars either.

[GitHub repo ↗](https://github.com/Folkbench/folkbench)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| cc-switch(farion1231/cc-switch) | Open source | 2025-08 | 152k stars, 10k forks | Pure switching tool, the de facto standard, no rankings for sale; Folkbench Switch is the watered-down version plus referral links |
| relayAPI (zzsting88/relayAPI) + HeWei AI ranking | Open source | 2026-03 | 5450 stars, 174 forks | Same play: relay-station recommendations with affiliate links. Started earlier. Detection code is closed source, and there's no switching client |
| OkkMax(okkmax.com / fanbidog/okkmax-web) | Web | 2026-07 | Open-source frontend, 201 stars | Focuses on model purity testing (thinking-signature verification) and user reputation. Declares it takes no money that affects rankings |
| Relay-station arena aiapipk.com | Web | 2024-12 | Not verified | One of the earliest batch, scored by the author's own call success rate and call volume |
| RouterHubs / APIRanking / llmtest.cn / hvoyai | Web | 2026 | Unverified (APIRanking claims to have tested 86 vendors) | Each focuses on invoices, authenticity rates or swap detection. Same niche |
| Grandet(hanzohanzhe/Grandet) | Open source | 2026-10 | 1143 stars, 125 forks | A Skill + MCP for agents covering 1800+ vendors' prices. Pure price comparison |
| ai-coding-welfare(panxunying) | Open source | 2026-08 | 1283 stars, 84 forks | Navigation for free / charity relay stations plus a one-click config script |
| relay-pulse(prehisle) | Open source | 2025-11 | 1109 stars | Self-hosted relay availability monitor |

Where users come from

- Inference **Bought social proof through star-farming, which then gets auto-reposted by aggregators and repost accounts** : about 230 downloads against 4312 stars, 0.49% forks, 0 days on GitHub's official trending, 11-07 lost -554 stars and -3 forks in one day. SkillsLLM, trendshift, LinkedIn post by Arthi R., an Instagram reel, explainx, kkzui's "3.8K-star relay live test leaderboard," and Zhihu's daily good-software roundup all cite the star count as the selling point

- **Confirmed** **Self-promotion and communities in Chinese developer circles** : 2 self-promo posts on V2026-10-16EX, thread t/1249363; the README links a QQ group 203286875 and a Telegram group

- **Confirmed** **SEO** : pinyin-slug landing pages /zhongzhuanzhan-ceping, /zhongzhuanzhan-na-ge-hao, /gpt-codex-zhongzhuanzhan-ceping, one leaderboard page per 48 models, in four languages, an English article on dev.to, and all README links carry utm tags

- **Confirmed** **Relay vendors self-onboarding** : a self-service submission page at /vendor/login; "listing request" issues appear in the repo

## Folkbench/folkbench: why it could run this way

> Research date 2026-11-06. Data sources: our GitHub database (repo_profile / repo_day, snapshot as of 2026-11-07), GitHub API, trendshift.io, folkbench.com pages, RDAP, Jina/Exa/Brave search, x-radar.

### Bottom line

Folkbench is **a traffic-driving repo for the relay-review site folkbench.com** , not a tool that grew through open-source users. Most of its 4000+ stars **don't look organic** : 6 version packages add up to only about 250 downloads, forks are just 0.5% of stars, it never made GitHub's official trending in a day, and 11-07 a day 554 stars and 3 forks were wiped out (a classic GitHub cleanup of fake stars). The star count is itself the growth tool: high star counts get auto-reposted by trendshift, SkillsLLM, LinkedIn/Instagram "trending repo" accounts, explainx, and tutorial sites, bringing exposure to the website. The site then makes money from **sponsor slots for relay stations + affiliate commissions (aff/partner codes)** .

### 1. What it is

- folkbench.com: lists AI API relay stations by model (48 models), with uptime, input/output prices and sample sizes. Also offers an "API quality check" (users test with their own key), a blog, and a relay-station directory, in Chinese, English, Japanese and Spanish.

- This repo: the desktop client Folkbench Switch (Tauri + Rust), which writes the chosen Base URL / Key into the config for Claude Code, Codex, Gemini CLI, Aider, OpenCode, Kimi CLI and others with one click (explainx 2026-10-18 summary).

- The first screen of the README is **a sponsor section** : Modelflare, CheapAI, IKunCode, FluxLane. Links redirect through folkbench.com/go/, and the comments keep the original partner= / aff= referral codes, plus a "reserved top sponsor slot" comment and a self-service "Become a sponsor → /vendor/login" onboarding page (README, read by 2026-11-06).

### 2. Timeline

| Timeline | Event | Sources |
|---|---|---|
| 2026-06-29 | folkbench.com domain registered | RDAP(verisign) |
| 2026-08-27 | "Last reviewed" date on the About page | folkbench.com/about |
| 2026-09-17 | Blog's first post, "The Grey-Market Ghosts of LLM Relay Stations" | folkbench.com/blog |
| 2026-10-02 | GitHub organization and repo created the same day (organization 09:44, repo 09:50) | GitHub API |
| 2026-10-07 | English SEO page "Which AI API Relay? October 2026" | Exa |
| 2026-10-11 | Switch beta.1/beta.2 released | Releases |
| 2026-10-16 | V2EX self-promo post "Tired of switching lines and remotely setting up Claude Code for friends, so I built a desktop tool called Folkbench Switch" | v2ex.com/t/1249363 |
| 2026-10-12~10-25 | Gaining 200~568 stars a day; 10-24 single day +568 (trendshift) / +10-25 561 (our database) | trendshift, repo_day |
| 2026-10-18 | explainx's write-up, citing about 1,664 stars | explainx.ai |
| From ¥2026-10-26 | Daily gains fell back to 50~90 | repo_day |
| 2026-11-05~06 | Star growth dropped to zero (trendshift +0/+0; our database +22, -10) | trendshift, repo_day |
| 2026-11-07 snapshot | **3758 stars (-554), 18 forks (-3)** | repo_day, GitHub API |

### 3. Why the star count can't be trusted (key point)

- **Downloads and stars are wildly mismatched** : the 6 Switch-version dmg/exe builds have roughly 230 downloads combined (beta.5 leads with Win 74, Mac arm 22, Mac x64 9; beta.6 has only 38). Against a peak of 4312 stars, that's about 1 downloads per 18 stars. Normal utility repos sit well above that. (GitHub Releases API, 2026-11-06)

- **Almost no forks** : at peak, 4322 stars came with only 21 forks (0.49%). Same-niche comparisons: zzsting88/relayAPI has 5450 stars and 174 forks, OkkMax's fanbidog/okkmax-web has 201 stars and 12 forks (6%), hanzohanzhe/Grandet has 1143 stars and 125 forks (11%). Our repo scoring tagged it as suspect: "almost no forks" and "new account, single repo surge."

- **On GitHub's official trending for 0 days** (our repo's trending_days_any=0); it has only appeared on trendshift's own Rust/TypeScript language charts (best Rust rank #2,10-22).

- **A one-time removal of 554 stars + 3 forks** (11-06→11-07) is the classic pattern of GitHub purging star-farming accounts. Meanwhile trendshift shows 11-05 and 11-06 star gains of 0, which looks like a starring service that has stopped.

- The organization has only 1 repos, and its sole main committer, the Lepusx333 account, was registered just 2026-10-10 and has 0 public repos.

- The star curve is step-shaped (10-12~14 three days +760, 10-18~20 three days +900, 10-22~24 three days +1300), dropping back to the tens between steps. That does not look like a single-peak decay from one HN/X post.

- Social reach is very weak: x-radar finds only 0 results for "folkbench" over the past 60 days, and no hot threads on HN, Reddit, or linux.do.

Conclusion: **real user numbers are very small** (Switch likely a few hundred people; website traffic unverified). It is highly likely the stars are "bought social proof" (we could not pull the stargazer list to check each account; the GitHub stargazers endpoint returned 404 from this environment).

### 4. Growth tactics (where users come from)

1. **Starfarming → auto-reposted by aggregator sites** (main engine, high confidence but not 100% confirmed): the high star count got it into trendshift's language charts, SkillsLLM ("4.3k★"), agentskillshub, and myfocalai trends. Then came LinkedIn "trending GitHub projects this week" repost accounts such as Arthi R. ("1k stars in under two weeks"), an Instagram reel ("Why developers love Folkbench"), an explainx blog, the Kuaizui Tech tutorial "3.8K-star relay station real-test ranking … step-by-step deployment", and a Zhihu column "Daily Good Software Picks 2026-10-19". The titles of nearly all these secondhand posts sell the star count.

2. **Self-promotion in Chinese developer communities** (confirmed): a self-promo post on V2EX 2026-10-16; the README links a QQ group (203286875) and a Telegram group; on GitHub, sharingcoke (a 2022-year-old account) submitted PRs to folkbench-blog and the main repo, suspected to be a team member.

3. **SEO** (confirmed): pinyin-slug landing pages /zhongzhuanzhan-ceping (24 relay stations), /zhongzhuanzhan-na-ge-hao ("which is good"), /gpt-codex-zhongzhuanzhan-ceping; one leaderboard page per model; four languages; the dev.to account folkbench publishes long English posts; all README links carry utm tags.

4. **Relay stations apply to be listed themselves** (confirmed): self-serve submission at /vendor/login; the repo has issues like "[Listing request] Twinkle Model — source relay, GPT 1% off". The entities being evaluated are themselves a source of traffic and revenue.

### 5. Where the money comes from

- Sponsorship: 4 sponsors in the README header, with a reserved "top sponsor slot."

- Affiliate commissions: /go/ redirects; original links carry partner=TDN1QM6KKYX9, aff=TkeF, aff=QgPu and other referral codes.

- **Conflict of interest in the rankings** : on the GPT 6.1 Sol chart at /zhongzhuanzhan-ceping, sponsor Modelflare ranks #1 with 93.75% availability, ahead of IKunCode (also a sponsor, 100% availability), ZNB.ai and ToolCode. Sponsor FluxLane, with only 18 samples, also makes the top 8 (2026-11-06 reads). So ranking is not purely based on measured results; at least price and commercial factors are mixed in.

- Revenue is not public (no TrustMRR, no founder revenue posts); the scale is unknown.

### 6. Competitors

| Name | Form | Launched | Scale | How it differs from Folkbench |
|---|---|---|---|---|
| farion1231/cc-switch | Open-source desktop switcher | 2025-08 | 152k stars, 10k forks | The de facto standard switcher, a pure tool that does not sell rankings. Folkbench Switch is a diluted version of it plus referral funneling. |
| zzsting88/relayAPI | Open-source README recommendations + Hewei AI leaderboard | 2026-03 | 5450 stars, 174 forks | Same "relay recommendations + affiliate" model, with detection code now closed-source; started earlier and its forks are normal. |
| OkkMax(okkmax.com / fanbidog/okkmax-web) | Website + open-source frontend | 2026-07 | 201 stars | Leads with "purity testing" (uses Anthropic thinking signatures to catch model swaps), states it takes no money that affects rankings, and has user word of mouth. Folkbench leads with availability + price + switcher. |
| Relay-station arena aiapipk.com | Website | 2024-12 (V2EX post) | Not verified | One of the earliest, scoring by the author's own call volume. |
| RouterHubs, APIRanking (86 sites), llmtest.cn real-rate ranking, hvoyai, oken.ai | Website | 2026 | Not verified | Each focuses on invoices, purity, or authenticity; same niche. |
| Chenking-maker/api-rank, fanbidog/ai-relay-rank | Open-source leaderboards | 2026-05 / 07 | 27 stars / 5 stars | The same "measured leaderboard" idea open-sourced, but it did not get stars. |
| panxunying/ai-coding-welfare | Open-source directories | 2026-08 | 1283 stars, 84 forks | Collections of free/public-good stations + one-click config scripts |
| hanzohanzhe/Grandet | Open-source Skill + MCP | 2026-10 | 1143 stars, 125 forks | Price comparison for agents across 1800+ official, inference, and relay prices |
| prehisle/relay-pulse | Open-source monitoring | 2025-11 | 1109 stars | Self-hosted relay availability monitor |
| qixing-jk/all-api-hub | Open-source browser extension | 2025-09 | 5049 stars | Manages balances, check-ins, and price comparison across multiple New-API accounts |

(Star counts from our GitHub repo and the GitHub API, 2026-11-06/07)

### 7. Differentiation

- Literally, the difference is that **it bundles a measured leaderboard with a desktop client that writes config directly into Claude Code/Codex** . Pick a result on the leaderboard and switch straight to it; other evaluation sites only give data, and cc-switch only switches. It also has four languages, a single-model leaderboard across 48 models, and many more SEO landing pages.

- But that is not why it took off: Switch's downloads are only in the hundreds; cc-switch already does switching best; the evaluation itself (availability + price) is less persuasive than OkkMax's "purity testing"; and the ranking mixes in sponsorship. It took off through **distribution** : packaging a referral site as an "open-source project" and trading stars for third-party reposts and a sense of credibility.

### 8. What to borrow

- The structure is "the website does business, the GitHub repo is the billboard": a sponsor slot at the top of the README, unified tracking via /go redirects, and utm tags across the whole funnel, using GitHub as a free, high-authority landing page.

- In a gray, high-commission category where users badly need to know "which one is reliable," evaluation-and-referral sites naturally get paid (sponsorships + affiliates), and the entities being evaluated come to apply for listing on their own.

- Long-tail Chinese SEO pages with pinyin slugs (relay station reviews / "which is good") + one page per model.

- Cautionary lesson: starfarming can buy a wave of aggregator reposts, but GitHub will purge it (here, 554 stars fell in a single day), and data where downloads and forks do not match is easy to see through.

### 9. Uncertain

- The starfarming is an inference based on downloads/forks/trending/one-time star removal; the stargazer account details were not obtained to verify it.

- Folkbench.com's real traffic, sponsorship revenue, and sponsorship pricing were not found.

- The operator's identity is unclear (explainx also says it found no named individual or company); whether sharingcoke is a team member is unverified.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.